← 返回

2026.10.02 | TB | Cloud | Fal.ai Customer View: Generative Media Infra

原标题：Fal.ai – Customer Angle – Scaling the Infrastructure Behind Generative Media

#cloud #ai-demand

Third Bridge · 2026-10-02

Analyst-led • 1 Oct 2026

Companies: GOOGL US

Christopher Barber (CB) Founding Design Engineer at Demi.ai

Dylan Koehler (DK) — Third Bridge Sector Analyst

Agenda

Key Insights

Introduction

DK [00:00:02]: Welcome to Third Bridge's Interview entitled Fal.ai – Customer Angle – Scaling the Infrastructure Behind Generative Media. I am Dylan Koehler, and I will be facilitating today's Interview with Mr Christopher Barber, former Head of Development at Silverside AI.

Chris, before we start today's Interview, please state I agree or I disagree to the following statement: You understand the definition of material non-public information and agree not to disclose any such information, or any other information which is confidential, during this Interview.

CB [00:00:27]: I agree.

DK [00:00:28]: Would you mind starting us off with a brief introduction to yourself and your experience working with Fal.ai?

CB [00:00:40]: I worked at Silverside as Head of Development. We are building both internal tools for AI in advertising. Silverside was born out of a larger advertising company. We always had our roots in advertising. We've built this internal tool to generate images, videos for our projects that we are doing with clients, and then we started selling that tool to clients themselves for them to use. That tool used Fal.ai for the most part. For some parts, we used other tools, but it did use Fal.ai for a majority of the media, API infrastructure, all these generative models and the long tail of generative models. That's my experience with Fal.ai, using them. We did a lot of content. Silverside did the Coca-Cola AI holiday commercial. We were very embedded in that world of AI and advertising, but yes.

Questions

00:02:07] Could you walk us through how a customer would use Fal.ai? Where does the company sit in the tech stack? What does it handle that a developer would otherwise need to build or manage itself?

CB [00:02:34]: Fal is essentially just like a generative model aggregator. They have image, video models, audio models, 3D generative models, all the models that you can think of. They just aggregate them on the platform and they allow you to programmatically call them, so it's just like an API aggregator. Developers use it, they can plug in Fal as like a library in their code where now, basically with one billing set-up, they can access thousands of models, which you couldn't necessarily do if you were connecting directly to the model providers themselves. You would have to create accounts at every single one of them and then set them up. From a developer perspective, it's just easier to set up just one account, so then you can access all of them and they handle some of the more complicated things, if you were to set up this, some of the infrastructure stuff like reliability, observability on whether or not your API calls are failing, your billing, how you actually bill your customers for how much media they're generating. There's some stuff that they not necessarily abstract away, but they make it easier for developers.

That's the core value proposition. The one other thing also just to mention is just like there's a new AI model every week that could potentially be useful to your company. It's just an insane rate of progress. Fal.ai generally makes it pretty easy because their developers will handle that. Then you can just switch a super easy string in for a new model and start using it right away vs having to implement it yourself or a more complex process. That's the core value, I would say.

00:05:00] You talked about Fal making it easy to switch between its models. As a customer, how much value comes from having access to thousands of models vs the handful you actually use in production? Does that breadth meaningfully increase customers' reliance on the company over time?

CB [00:05:36]: Yes, it's interesting. Just from my experience at Silverside, we did use such a breadth of models. Some of these generative models are better at one thing than the other. Sometimes we use a couple of different video models in the same video just because one is better at this type of shot where a person is running vs a shot where, I don't know, it's more like a motion graphics shot or maybe it's a person jumping instead of a person running and one of the models is better at it. There's some of those quirks where it's nice to have a bunch of different models that you can access. I think it's more about the long tail of just like they have three model generators in here too. If you want to do more stuff, if by chance a project comes up where you might need some of that, it can be good. At the same time, if there is a model that you're using a lot and Fal has that 10% margin on it, sometimes it could make sense to implement.

00:07:25] You talked about how Fal is an aggregator and provides access to a lot of different models, but if you're doing a lot of work with just one specific model, it might make more sense to go direct. How much engineering time does it actually save to use Fal for several models rather than going to those models directly? Does it save a couple of hours per week? Does it potentially circumvent the need for dedicated engineering headcount to build the models out?

CB [00:08:35]: That's a good question. I think it depends. If you were to implement a new model every week, I would probably say, yes, it's at least a couple of hours of an engineer's time. With coding agents, it's really quick to change things out now, but you have to test it. If you have a lot of users, you'd want to go through a whole testing process, you'd have to check the reliability of going through the provider, whereas you might have to hook it into your logs so you know, okay, when do errors happen? It just increases the complexity of your system if you're connecting to, let's say, 10 different providers vs just one provider of Fal. It's definitely saving you some time to just do Fal. Frankly, their margins aren't that much higher than the actual cost of the model, so it's not crazy to just go through Fal. I'm sure a lot of companies just do Fal completely. It definitely does save some time, I would say, at least a couple of hours per week.

00:09:59] For customers using several different image and video models, how would you expect workloads to be split between an aggregator such as Fal and direct integrations with model providers? You used the example of using Fal regularly, but using Nano Banana a lot, so having a direct integration for that specific model. What does that split look like?

CB [00:10:37]: I think it would probably be like 60% Fal, 40% others. Just like if you do have the time, I think it can cause problems down the line.

00:11:28] Do you expect Fal's share of customer inference spend to increase, remain stable or decline as people do direct integrations over the next 1-2 years?

CB [00:11:50]: I would imagine everything would increase just in terms of media inference spend. Maybe Fal's share of it is not going to increase a ton, but the media inference spend of a given company is going to because every company that we worked with was essentially trying to figure out how to use AI more in their, whatever it was, e-commerce, actual commercials, ads on Facebook. All of them were just trying to figure it out. A lot of them weren't there yet. We were like the starting point of teaching their creatives how to use these tools. I think it's only going to increase in spend in general. Maybe not their market share of that vs going to the providers themselves, but in general, I think media inference spend is going to increase.

00:13:01] If you're saying Fal's share doesn't increase, so its split of spend stays around 60%, what do you think could drive the underlying spend growth the most? Are more companies adopting generative media? Are existing customers generating more content? If that underlying spend is going to grow, what's driving that?

CB [00:13:48]: It's a bunch of things. I think one thing that we saw with our customers a lot was just, one, they were just learning how to use it and they're figuring out use cases in their company where it could be useful. It's kind of insane, if you compare the cost of traditional advertising and film shoots, even a photo shoot, the costs are pretty astronomical. You could look at something like a couple of hundred to a couple of thousand dollars per photo at a photo shoot, for example, whereas it costs like USD 0.05 to generate an image or USD 0.10, maybe USD 0.15. They were already spending that much on photo shoots, so they have a lot of room to spend, these companies. The thing that you could do with AI is version your assets. We did this sometimes, you could still shoot it in person, and then we could version your assets, I don't want to say infinitely, but we can version them like, let's say, you did a photo shoot in the spring and it was spring themed but you wanted to reuse that photo shoot in the winter, you wanted to make it more winter feeling for a clothing store, for example.

Those are the kinds of things that you can do. You can personalise it to a specific maybe time or, "It's Mother's Day, can you personalise it to a Mother's Day shoot?" You can do things like that. Beyond that, I think a lot of the companies hadn't even tapped into demographic variations in content. You could direct your content depending on who you're serving it to, like on ads, not just by caption, which is like, okay, people are doing that on Facebook ads all the time, but the actual visuals and really changing that per demographic that you're trying to sell to. I think a lot of that hasn't been unlocked. That's where I see the most growth of just like if these models are this cheap, like it's USD 0.10 per image, why would you not do 100 variations of the image if you can guarantee quality? Which obviously, that's a problem, but the models are getting better. If the quality gets there, then it just makes sense for a business to do 100 variations of a project.

In my eyes, okay, inference spend is only going to increase. There's also the fact that, right now, AI agents, people are exploring a lot of just the agents themselves generating content and generating variations of content. When you're doing stuff with agents, they can do way more variations than a human might do, hundreds and thousands of variations. You could imagine you have sort of like a feedback loop of just like you do thousands of variations, you post them on these ad platforms like Meta or something, then you see how they do and you create more content of the type that's performing well and so you do thousands more variations of those. It's like an ad evolution network of just like a feedback loop of getting better and better. I just think inference spend is just going to increase so much just from those factors. There's probably even more factors that might increase spend, but that's what I've been seeing.

DK [00:18:24]: It seems you're saying that, given how early customers still are in these areas, the asset versioning that you were talking about – the personalisation and demographic targeting – has substantial runway for generative media inference spend to continue to grow over the next several years. Is that right?

CB [00:18:48]: Definitely. There's Fortune 500 companies that just haven't really tried it in their advertising yet or haven't really pushed it at all. That's what I would say there.

00:19:03] In a September 2026 article, Fal was reported to have reached roughly USD 800m in annualised revenue, doubling from around USD 400m in March 2026. Based on the demand that you saw for image and video generation, how sustainable does that pace of growth look over the next 1-2 years? Is that the pace of growth you expect to continue to see for broader generative media inference spend as well?

CB [00:19:55]: Yes. That's crazy how fast they're growing. This is a rocket ship. I think it could stay on track. It's not insane to me that over the next year, companies start using it way more. I feel like image and video models haven't quite been adopted as much as just LLMs in the workspace. I just feel like they're just more getting there of just people are integrating them, like image workflows into design, when you're designing something like web design or graphic design or posters, videos. It's just seeping into everywhere. I think it's totally possible for the growth to just stay on that track for another year just because I just think there's so much for companies to do in that space, like using generative models in parts of their business. Granted, I would say, in advertising specifically, it depends on the company, but there is obviously a lot of backlash of using AI in advertising.

Some companies are still shying away from it and scared of it, which granted a lot of their consumers would really get angry if they did. I would say vast majority of consumers aren't really noticing it or caring about it that much. That's the only consideration of maybe some companies just won't touch it from that, but I still think the companies that do want to play around with media inference and seeing how they can fit it into, they're just going to continue. I saw this at Silverside before I left, we were just getting more and more clients. It's like everyone is interested in what they can do in this space just because it's one of those clear ways that even a marketing exec or any exec in their company can be like, "We're actually saving money because we used to spend this much on advertising, on photo shoots and now we save that money with doing it in AI." It's a clear thing you can point to of like, "We're saving money with AI," which oftentimes hasn't been happening because people have been spending a lot on AI tokens for like software engineers and for other parts of their business.

00:23:16] You mentioned that you think maintaining the current growth trajectory for another year feels possible for Fal. If the company's annualised revenue doubled again, it would go from USD 800m to USD 1.6bn. Based on the adoption you were seeing, does that seem achievable, or do you expect growth to moderate between now and then, but stay at a very impressive rate?

CB [00:24:10]: I can't say for sure, but I think it makes sense to me. I think that it would continue on this trend just because I think spend is going up, these models are getting better and they're also getting cheaper, which it feels paradoxical. It's like, okay, then Fal is going to make less money, but when these models get cheaper, it just unlocks more capabilities because you're like, "Oh, I can use it for this. I can use it for this now because it's so cheap that I could make 100 variations, no problem, of videos now." I think there's still a lot of room for growth, honestly. We saw it with coding models or with Anthropic, their revenue 10x this year from last year. I think it's totally in the realm of possibility.

00:25:19] You mentioned that if generation is getting cheaper, there's more room for growth, so people are doing more with it. If falling inference costs could drive more substantial usage, how quickly would you expect generation volumes for a typical Fal customer to grow over the next 12 months? Would it be in line with the rate of revenue growth that we were just talking about – 100% YoY? Or do you expect those volumes to increase 50%-plus or 25-50%?

CB [00:26:47]: It's hard for me to say exactly. Do you have an idea of what a typical Fal.ai customer is? I don't know their exact breakdown. From my point of view, yes, I don't think 100% growth… I know it sounds crazy, it's like, "Okay, that's a lot," but that doesn't sound like that insane to me. If a bunch of companies like Silverside, these new-age advertising companies or even the advertising companies themselves, we talk to a few advertising companies that were coming to us like we are consulting them because they don't necessarily know how to use AI and they want to use it more. All of these companies are making internal tools. I would imagine Fal is just one of the best, maybe there's a couple of competitors, I don't know if they're using other competitors. I would imagine if these advertising companies are spending more, companies are more willing to, whatever it is, build an internal tool for advertising. We saw this with a couple of our clients. We were doing stuff for them, but on the flip side, they were pretty big companies so they were doing their own internal tools for generating content.

I could see it increasing. I also think, like I said, the agent stuff will be a big factor into it. I know Fal, they do a lot of experimentation, which I think is one reason why developers really like them is they're always experimenting with these new models, kind of pushing them. With agents and stuff, they like to push them. How far can you go? I think AI agents using Fal to generate tons of content could increase their media spend as well, could increase inference spend. I think 100% growth, I don't want to sound too bullish on them, but yes, that does not sound crazy to me.

DK [00:29:30]: I think it's difficult to say what a standard Fal customer looks like. The company has partnerships with Adobe, Canva and Shopify. I'm sure it has a lot of specialised generative AI start-ups and marketing platforms as well.

00:29:30] You mentioned agents as another potential major driver. How would an agent use Fal.ai differently from a human developer? Why might that result in higher inference volumes?

CB [00:30:29]: How an agent would use it differently is it might use the Fal API. First of all, I think they built like a harness for an AI agent, basically, an agent can use all the different models from Fal and essentially, they have guides, skill files for every single model or every type of model, probably not every single one, which is nice to have. The AI agent knows exactly how to use this specific image model or this specific video model potentially better than a human could do just because if you… The generation process, I will say, just from Silverside, it's a lot of generations. For one video project, we might do thousands of image generations to storyboard it out. Those thousands of image generations, you could imagine we were doing them semi-manually, where maybe, this sounds counterintuitive to some people, but you'd have ChatGPT write the image prompt or help write the image prompt and then generate images from there.

You might storyboard out a project, let's say, there's 20 shots and so you're storyboarding out. It depends on the shot, but it might take you, I don't know, 20-100 generations per shot just to get that right, like image keyframe that you use for video generation. Then there's video generation itself, which is even higher inference spend, and you might do tens to hundred video generations for each shot. You're talking about a lot of generations here, and you're talking about things that agents are really good at, which is like prompt variations, just testing out a bunch of different models. What I was saying where different models are good for different things, there are so many different models now, it's hard for a human to go in and test. How do we test 10 different models? It's a bit of a manual process and we probably wouldn't want to do that, but an agent could go in and be like, "Okay, I tested it on these 10 different models. This style of video that we're doing for this client looks the best on ChatGPT image generation," for example. "Let's use this model." Then the agent goes in, so it's like literally hundreds of variations.

The human picks out the best of those shots. Now you can imagine if we're using these agents correctly, and they're helping us storyboard and come up with the content where we can use our creatives to just direct the models on where they should try out things, pick out the best images and the best videos and combine them. You could imagine an agent is just doing much more variations, much more content. I would say, every time we signed a contract with a client, it would be like there would be more variations involved. At the beginning, we were just like, "Okay, we'll do one video for you." Then these next companies, it was like, "We'll do one video and we'll vary it five times for different locations." Then people wanted variations on all the images that you were creating for them. Companies want this variation, agents are how you get there. I think, yes, using agents in general will just increase your inference spend, but still less than what they might spend on traditional advertising, much less.

00:35:31] If agents allow a company to go from concurrent projects while also generating more variations per project, how much higher do you think that total generation volume could be? If we could see serious growth in volumes and falling inference pricing to offset it, does the actual dollar spend grow at a lower rate than the generation? Do you think the pricing declines make spend growth meaningfully slower than volume growth?

CB [00:37:03]: It's always hard to say how much cheaper image and video models will get. There's all these Chinese models, for example, that are much cheaper that are oftentimes open source. Seedance isn't open source, but a lot of the other ones are. In some ways, yes, there are models that are getting much cheaper, like for image generation, for example. I think what I'm seeing is inference spend is outpacing the rate at which image generations get cheaper. I feel like image generations are already at USD 0.10 per image. Video generation, there's definitely a lot of room to improve. You're talking sometimes USD 0.50 to a couple of dollars per video, which is pretty expensive and can add up if you're doing hundreds of videos, still bounds cheaper than a traditional advertising shoot, which might cost USD 400,000-1m. You're still saving a lot of money there. I would imagine with agents, that increase, I don't have numbers, but maybe you could do, human creative, one of our creatives would maybe do… I don't know, sometimes for storyboarding, it would take even one day to do a couple of shots on the storyboard that I was talking about, so maybe one or two or three. That's maybe tens to maybe a couple of hundred generations in a day.

It's hard to imagine. An agent could do all 20 easily. If you were willing to pay it, in one day, an agent could do all 20 shots on the storyboard and do 2,000 or a couple of thousand generations in a day. If you had the right set-up, I'm not saying everyone has this set-up that's going to be perfect and they can storyboard, I think it takes time to build that set-up, but what these companies like Silverside and some of these other new media companies and some of these larger companies that have internal tools, they're figuring it out. They're learning how to make these workflows that might work. The amount of inference that you would use is just so much more, I think, than the pace of inference getting cheaper, which it is getting meaningfully cheaper, but I still think the amount of variations that someone can make with an agent is just so much more.

00:40:40] You mentioned that the best image or video model could change every few weeks. How important is that fragmentation to what Fal is and what it offers to the customer? The company's service seems important in a world where things are changing rapidly, but if there are 2-3 models that eventually become clearly superior, how much does that weaken the need for an aggregator such as Fal?

CB [00:41:25]: I think it's super important. I can't really see a world where only two or three image models. There's so many companies that are trying to do it. There's already just so many models out there. There's the open-source models that Fal.ai themselves, they can put some of their engineers on and see if they can make it faster or cheaper or better, and those could be used for specific use cases. There's some of that, which is nice that Fal does. I think in general, it is important. I think one of the annoying bits or one of the reasons why you would use Fal is just because every couple of weeks, there is a model that you have to go and implement and so it's just saving you that time. Beyond that, you could argue with LLMs themselves, there is two or three. It's like Anthropic, ChatGPT and Google, but for some reason, people are still using OpenRouter, and it's just because it makes it easy to set up billing and all this infrastructure on logging. When a new model comes out, OpenRouter has it handled, same thing with Fal.

If you're saying, oh, if there's only going to be two or three companies, even the new models from those companies, that takes time to plug into your system if you didn't have something like Fal. I think that would be where my head is at. It's still going to be useful regardless, but I don't think there's only going to be 2-3 image models, but people would still use it regardless, even if it wasn't. It's important that there's many models every couple of weeks, but I wouldn't say it drives the whole reason why someone would use Fal.

00:44:04] Fal has differentiated itself as a media-specific infrastructure and inference optimisation service. How noticeable are differences in latency, reliability or output and performance between the company and other ways of accessing the same model through competitors? Are those differences large enough to influence purchasing decisions for the customers?

CB [00:44:55]: Not a ton. I don't have their numbers, but I would imagine they're all pretty similar. I also think a lot of the usage is just going through Fal to Google or something, so it's not necessarily Fal optimising inference. They optimise the inference on open-source models, but I would imagine a lot of the usage is on non-open-source models, where they're basically just routing the API calls through Fal. I don't think there's a ton of variation between Fal and their competitors. I would say the biggest differentiator for Fal is just their developer experience is what I would say. I think developers, at least I appreciate that, yes, they are trying to do these inference efficiencies and they are exploring these different models, pushing open-source models to their limits and I think they have their own models now. They're also trying to do some of that as well, more vertically integrate. To answer your question, not a ton of difference between Fal and another company.

00:46:23] In our pre-Interview conversation, we talked about it not being particularly difficult to switch away from Fal, but there are reasons why you would stay, such as billing, usage tracking and other integrations that might create additional friction. What would a competitor actually need to offer to make someone switch away from Fal? Is it even possible for another platform to be 10-15% cheaper with similar model coverage and reliability? Would that be enough for a customer to switch? Is the developer experience that you were talking about valuable enough that it justifies a pricing premium for Fal?

CB [00:47:32]: I think it would be really hard for another provider to do 10-15% lower. I just think they would be losing money on every call. They could. If they raised a lot of money, maybe they could try to, I don't know, do what Uber did and just make things cheaper just for the time being and then make it more expensive. They can make it a little bit cheaper, but it does have to be substantially cheaper for someone to switch to it because it's like, yes, it is a friction, you're spending a lot of developer time, which frankly, unless you're doing an insane amount of inference spend, the developer time is probably much more expensive than what you're spending on these models. I don't know. It depends how much you're spending, of course. I think there is substantial friction. I think where there probably is room for the competitor growth is just, who do they go to first when there's new companies or companies are making their internal tool and they want to use an inference provider? Maybe they'll go with one of their competitors rather than Fal.

I think it's generally pretty sticky because it's like you already know, okay, Fal has good developer experience, they're reliable, they've got good customer service, you can go to their Docs platform and just look at how their API works and their docs are laid out well and organised and they've got good branding. I don't know. I think there's a couple of things that might draw you to Fal just to stay. I don't think 5% is enough for you to switch.

00:49:47] If existing customers are relatively sticky, then it seems that the real battle for Fal is the new workloads rather than displacement. As this market grows, what's going to determine whether a new developer chooses the company over competitors such as Together.ai, Fireworks.ai or Baseten?

CB [00:50:42]: It's interesting. I think Fal has carved out this space of media inference. Fireworks, for example, does a ton of LLM stuff, too. I know Baseten does a ton of LLM stuff. I don't know how much media inference focus they have there. I'm not sure about Together.ai. I think, yes, they have an interesting thing where they have carved out that space. If you're thinking about media inference, you might want to go there. I just think they have pretty solid online presence and branding in that regard of why someone might choose them of just like if you ask ChatGPT or if you search up online, best inference provider, they're going to come up. Beyond that is also I feel like they're exploring with the agent stuff a little bit more than other companies or maybe not more than other companies, but at least they're really good at showing that they're doing it. They'll do blog posts on, "Oh, we tried this out, we created a new Fal agent," and they have some of that AI agent stuff that I haven't seen other companies doing as much of a push for that.

00:52:37] You mentioned the 10% margin that Fal is getting when people are inferencing on its platform. Based on the current product and developer experience, how much of a premium do you think a typical customer would be willing to pay before it becomes worthwhile to bypass the company? If it's charging a 10% premium above a direct integration, can it push that higher than 10%? Could it get 15% or even 20%?

CB [00:53:42]: That's a good question. I think as models get cheaper, there might be more willingness to pay some of these premiums just because it's already so cheap. I do think there is a point where potentially, if some company did the calculation, they're like, "Guys, if we switch from Fal to Baseten, we'd save however much per year." I think there could be that. One of the good things about it is that it's pretty confusing. Pricing for these models is relatively confusing. I had to get in the weeds of it in the spring of this year, and essentially, video models can sometimes price per megapixel per second or sometimes they just charge you for the entire video, for a five-second video. Same thing for image models, sometimes they charge you per megapixel, sometimes they charge you just for generating the image themselves. If you do it with OpenAI, there's a high mode, a medium mode and a low mode and they all are priced differently. This, I think, works in their favour because it's like this pricing is already confusing. As a developer, I don't really mind, I wouldn't maybe even notice, I wouldn't be checking Google's provider website and seeing how much they're charging.

You might not notice just because the pricing is pretty confusing. I know it seems like a weird reason, but I think it's definitely a factor of where I don't know what their exact margin is, I never calculated it, but I think that works in their favour. I think as a developer, you're probably not thinking about it. If there is someone on the executive side that's like, "Wait, we could save this much money," I think they might, but again, I think the cost of the inference is probably so much smaller than the cost of, let's say, paying their employees to switch to a different provider. I think it'd be hard to see. I think they could increase their prices a little bit. I also don't know if they need to. I think they could still make a lot of money in staying at a smaller margin.

00:56:55] We've talked about agents so far, but Fal is expanding beyond the basic model access with Fal Agent, looking to select and change models automatically. Do you view owning the agent orchestration layer as critical for the company going forwards?

CB [00:57:36]: I think it can be valuable. I think if it's good, if they spend a lot of engineering power on making this agent really good, if you give it a task in the media space like, "Make this commercial," and it comes out with a really good output, yes, there's definitely value. We see it with coding models, there's all these companies that have created their own coding harnesses. Cognition, they just raised USD 2bn, Factory, also. They're basically just coding harnesses. It's the same thing there. This is just like a creative harness, not a coding harness. I think there is a lot of value there. It's hard because a coding harness or a creative harness, some other person could make a similar type of thing pretty easily, but if you invest time in and you make it really good, then I think there's definitely value there just from what I've seen with other companies.

DK [00:59:51]: I think that's a great place to end things for today. I want to thank you for your fantastic input over the last hour. I really appreciate all of the time and help today. Thank you so much for that, and thank you to our clients for listening and reading Third Bridge's Interview on Fal.ai today.

CB [01:00:11]: Awesome. Thank you. I appreciate it.

Transcription ends at 01:00:14 of the recorded material